Technical Summary: LadderMan

Problem Statement

Enabling humanoid robots to operate in human-centered environments requires robust ladder climbing capabilities. However, this task presents significant challenges compared to flat-ground locomotion or stair climbing. Ladder climbing demands precise, sustained coordination between upper and lower body contacts on sparse, discontinuous, and narrow footholds and handholds. Small errors in perception or control can lead to catastrophic failure. Furthermore, real-world ladders vary significantly in inclination, rung spacing, and structural design, making generalization difficult. Existing approaches often rely on model-based control requiring accurate environment modeling and task-specific tuning, or reinforcement learning (RL) methods that produce lower-body-dominated behaviors failing to exploit whole-body coordination. Additionally, few methods address on-ladder manipulation, where the robot must maintain balance while performing tasks.

Methodology

The authors propose LadderMan, a unified system for perceptive humanoid ladder climbing and on-ladder manipulation. The system employs a scalable two-stage learning pipeline and a dual-agent framework for manipulation.

1. Learning Multiple Expert Policies (Stage 1)

To avoid the need for large motion datasets covering diverse geometries, the authors introduce hybrid motion tracking.

• Single Reference Motion: The system learns multiple state-based expert policies ($\pi_{\phi,z}^{expert}$) corresponding to different ladder configurations (inclination $\phi$, rung spacing $z$) from a single reference motion.
• Asymmetric Tracking: The tracking reward is partitioned into upper and lower body components. The lower body is tracked strictly to ensure stable climbing behaviors, while the upper body is tracked with relaxed constraints to allow adaptation to specific ladder geometries.
• Task-Driven Adaptation: The system incorporates ladder-centric contact rewards and climbing rewards. Contact positions are automatically generated based on the target ladder geometry, enabling the policy to adapt contact locations without manual annotation for every configuration.

2. Unified Visuomotor Policy Distillation (Stage 2)

The multiple expert policies are distilled into a single depth-based visuomotor policy ($\pi_{visual}$) using hybrid imitation and reinforcement learning.

• Hybrid Objective: The policy is optimized using a combination of a PPO-based RL objective (using the same rewards as the experts) and a KL-divergence term to encourage imitation of expert behaviors. The weight of the imitation term is annealed during training, allowing a transition from imitation-driven initialization to task-driven optimization.
• Inputs: The unified policy takes proprioception, depth observations, and a binary climbing direction command as input.

3. Bridging the Sim-to-Real Gap

To enable zero-shot transfer to real-world hardware, the authors address the sim-to-real gap in depth perception, which is critical for thin ladder rungs.

• Vision Foundation Model (VFM): Instead of relying on extensive depth randomization, the system uses a pretrained VFM (Fast-FoundationStereo) to process real-world stereo grayscale images into depth maps. This provides cleaner, geometrically consistent depth observations.
• Rung-Focused Masking (RFM): During training, non-rung regions are masked with a probability of $p=0.1$ to force the policy to focus on task-relevant climbing structures rather than background geometry or specific ladder supports.

4. On-Ladder Manipulation via Dual-Agent Learning

For manipulation tasks while on the ladder, the authors employ a dual-agent formulation to decouple stabilization from manipulation.

• Lower-Body Agent ($\pi_l$): Responsible for maintaining stable balance and ladder contacts.
• Upper-Body Agent ($\pi_u$): Tracks teleoperated manipulation targets (arm, wrist, waist poses).
• Decoupling: By using separate actor-critic networks, the system reduces interference between the stabilization and manipulation objectives, enabling stable whole-body teleoperation on the ladder.

Key Contributions

1. Unified Perceptive System: LadderMan enables robust zero-shot sim-to-real ladder climbing across diverse geometries (inclination, rung spacing, structure) using a single reference motion and depth-based visuomotor control.
2. Hybrid Motion Tracking: A novel method to learn multiple expert climbing policies from a single reference motion by combining relaxed upper-body tracking with task-driven contact adaptation, eliminating the need for large motion datasets.
3. Sim-to-Real Perception: The integration of a Vision Foundation Model and rung-focused masking to bridge the depth perception gap, allowing deployment on real-world hardware without extensive manual tuning or hardware modifications.
4. Stable On-Ladder Manipulation: A dual-agent learning framework that successfully enables complex manipulation tasks (e.g., adjusting paintings, replacing light bulbs) while maintaining balance on a ladder, outperforming standard whole-body teleoperation policies.

Experimental Results

The system was evaluated on a 1.3m, 29-DoF Unitree G1 humanoid using an Intel RealSense D435i camera and NVIDIA Jetson Orin.

• Real-World Performance: LadderMan successfully climbed up and down three diverse real-world ladders (varying in angle, spacing, and material) in a zero-shot manner. It achieved a climbing speed of 3.4 seconds per rung, comparable to human subjects (3.2 s/rung).
• Manipulation: The system successfully performed painting adjustment, light bulb replacement, and box handover tasks while maintaining balance.
• Generalization: In simulation, the policy achieved >95% success rates on ladders with rung spacings between 24-28cm and inclinations between 55-65 degrees. It maintained ~57% success even at extreme boundary configurations.
• Ablation Studies:
  • Expert Learning: Standard motion tracking failed on geometries different from the reference, whereas hybrid motion tracking succeeded across all tested configurations.
  • Sim-to-Real: Removing the VFM, rung-focused masking, or the RL component in the distillation stage significantly reduced real-world success rates, with some variants failing completely.
  • Manipulation: The dual-agent policy maintained stable balance during manipulation, whereas a state-of-the-art baseline (TWIST2) frequently lost balance.

Significance and Limitations

The paper claims that LadderMan represents a key milestone for deploying humanoids in industrial environments (construction, maintenance) by solving the challenging problem of perceptive ladder climbing and manipulation without hardware modifications.

Limitations acknowledged by the authors:

• The climbing policy is currently evaluated on inclined ladders up to 75 degrees and does not yet extend to vertical ladders, which require different strategies and higher contact forces.
• Manipulation capabilities are currently limited by the absence of dexterous end-effectors on the current hardware.

The authors state that all training and inference code, along with deployable models, will be open-sourced.